Different architectures to develop repetitive

controllers
Vı́ctor Sanz i López ∗ Ramon Costa-Castelló ∗∗
German A. Ramos ∗∗∗
∗
Institut de Robòtica i Informàtica Industrial, CSIC-UPC Llorens i
Artigas 4-6, 08028 Barcelona, Spain. Email: vsanz@iri.upc.edu
∗∗
Universitat Politècnica de Catalunya; Barcelona, Catalunya, Spain.
Email: ramon.costa@upc.edu
∗∗∗
Universidad Nacional de Colombia; Bogotá, Colombia. Email:
garamosf@unal.edu.co

Abstract: In this work several architectures to implement repetitive controllers are compared.
A complete analytical analysis is performed for these different architectures and a simulation
example with a power converter is included.

Keywords: Repetitive Control, Disturbance observer, Youla parametrization, inverter control

1. INTRODUCTION example is presented and finally section 8 provides some

discussion about the different approaches.

Repetitive Control (RC) (Wang et al., 2009; Longman, 2. INTERNAL MODEL

2010; Chen et al., 2008; Ahn et al., 2007) is an Internal
Model Principle (IMP) (Francis and Wonham, 1976;
Costa-Castelló et al., 2012) based control technique. RC
it is specially suited for systems subject to periodical
references and disturbances (Ramos et al., 2013). During
last years this technique has been applied in different fields
like power electronics (Ramos et al., 2013) or mechatronic
systems (Park et al., 2005) among others.
The idea behind RC is to include inside the control-
loop the generator of a periodical signal, which is a high
order dynamic systems. Applying conventional stabilizing
techniques to this type of systems would imply obtaining
very high order controllers which would entail huge
computational resources to implement these controllers.
Due to this, RC uses specific architectures and anti-windup
techniques (Ramos and Costa-Castelló, 2013; Wang, 2016)
which allow to take profit from the steady-state nice
properties of RC and obtain implementation structures
which allow reducing computational requirements to a
minimum.
Since the seminal work by Inoue et al. (1982), which intro-
duced the plug-in architecture for RC, many architectures
have been proposed, both in the input-output (Chen and
Tomizuka, 2014, 2015) and the state-space (Wu et al.,
2014) formalisms. In this work most relevant input-output
based architectures are reviewed and compared. The com-
parison is illustrated in the case of a power inverter.
This paper is organized as follows: section 2 describes
most relevant concepts about internal models used in
RC, in section 3 the series architecture is described, in
section 4 the plug-in structure is introduced, in section 5
a disturbance observer approach is used, in section 6 the
controller parametrization is used, in section 7 a numerical
I(z)
E(z) + Y (z)
σ · W (z) H(z)
+
Figure 1. Internal model used in Repetitive Control.
The internal model is the core element of a RC system
and it is composed by an element which can generate a
periodic signal. Figure 2 shows the structure of a generic
internal model used in RC:
σH(z)W (z)
I(z) = ,
1 − σH(z)W (z)
where σ = {−1, 1}, W (z) is the time delay function, and
H(z) is a low-pass filter. As an example, for σ = 1,
W (z) = z −N and H(z) = 1 a N-periodic generator,
N
I(z) = zN1−1 , is obtained and for σ = −1, W (z) = z − 2
and H(z) = 1, the odd-harmonic generator (Griñó and
Costa-Castelló, 2005) is obtained I(z) = N−1 .
z 2 +1
For those internal models obtained with H(z) = 1,
the internal model, I(z), introduces infinite gain at the
signal frequency and all its harmonics. This high gain at
high frequencies might be a problem in the presence of
uncertainty. To reduce this gain a low-pass filter, H(z), is
usually used. A null-phase low pass filter is usually used
(Griñó and Costa-Castelló, 2005; Escobar et al., 2014).
 I(z) this internal controller is to guarantee closed-loop stability
and robustness to the control system without the internal
+ model. Later, the internal model I(z), and the stabilizing
R(z) + Y (z)
σ · W (z) H(z) Gc (z) G(z) controller, Gx (z), are plugged in to the previous closed-
− + loop system.
In this case the closed-loop transfer function can be
Figure 2. RC : Series approach. constructed in terms of the closed-loop transfer function
without the internal model:
3. SERIES APPROACH
1 Gc (z)G(z)
So (z) = , To (z) =
The first approach to RC consists in placing the internal 1 + Gc (z)G(z) 1 + Gc (z)G(z)
model in series connection with the plant. Additionally a and a modifying term:
stabilizing controller needs to be introduced to guarantee
closed-loop stability (Figure 2). In this case the controller 1 − σW (z)H(z)
becomes: SM od (z) = .
1 − σW (z)H(z) (1 − Gx (z)To (z))
C(z) = I(z)Gc (z)
So the closed-loop transfer functions are:
the complementary sensitivity and sensitivity function
become:
S(z) = So (z)SM od (z)
I(z)Gc (z)G(z) (1 − σW (z)H(z) (1 − Gx (z))) To (z)
T (z) = T (z) = .
1 + I(z)Gc (z)G(z) 1 − σW (z)H(z) (1 − Gx (z)To (z))
1 For minimum-phase plants 1 the most popular form of for
S(z) = .
1 + I(z)Gc (z)G(z) the stabilizing controller is:
Usually, for minimum-phase plants kr
Gx (z) = . (2)
kr To (z)
Gc (z) = (1)
G(z) With this selection the closed-loop function becomes:
and :
1 − σW (z)H(z)
kr σW (z)H(z) S(z) = So (z)
T (z) = 1 + (kr − 1) σW (z)H(z)
1 + (kr − 1)σW (z)H(z) (To (z) − σW (z)H(z) (To (z) − kr ))
1 − σW (z)H(z) T (z) = .
S(z) = . 1 + (kr − 1) σW (z)H(z)
1 + (kr − 1)σW (z)H(z)
As expected, S(z) has in the numerator the denominator of The following two conditions guarantee closed-loop stabil-
I(z). For non-minimum phase plants it is necessary to use ity (Ramos et al., 2013):
phase cancellation techniques (Ye et al., 2009; Tomizuka, (1) To (z) must be stable (Gc (z) can be designed to fullfil
1987) when designing Gc (z). it)
In case of systems with multiplicative uncertainty: G(z) = (2) kW (z)H(z) (1 − kr ) k∞ < 1 (kr can be selected
Gn (1 + Wum (z)∆(z)) (Sánchez-Peña and Sznaier, 1998), appropriately )
the robust stability condition is: Even thought these conditions are only sufficient it has




m

Wu (z) kr σW (z)H(z)


. been proved that they are close to the necessary ones in

1 + (kr − 1)σW (z)H(z)
practice (Songschon and Longman, 2003).
∞
It is important to emphasize that in (2) the inversion of
4. PLUG-IN APPROACH To (z) is required while in (1) it is required the inversion
of G(z), as To (z) is a closed-loop system its uncertainty
C(z) should be less than that of G(z). Additionally, it is
I(z)
important to visualize that the sensitivity function in the
+ series approach and the plug-in one are the same except
σW (z) H(z) Gx (z) the So (z) term.
+ +
R(z) + + Y (z) In case of multiplicative uncertainty the robust stability
Gc (z) G(z)
− condition

becomes:


m (To (z) − σW (z)H(z) (To (z) − kr ))


Wu (z)
< 1.

1 + (kr − 1) σW (z)H(z)

∞
Figure 3. RC : Plug-in approach. This condition is quite similar to the one obtained in the
series approach but it contains To (z). This term can be
The most popular approach in RC is the plug-in architec- shaped using Gc (z), so it is simpler to fulfill this constrain
ture (Figure 3). This architecture introduces the internal than the one obtained in the series approach.
model to a previously existing control system defined by an 1 For nonminimum phase plants a phase cancellation approach is
internal controller Gc (z) and the plant G(z). The goal of usually used
 5. DISTURBANCE REJECTION APPROACH degrees of freedom they might increase the controller
complexity.
C(z)
+ + 6. YOULA PARAMETRIZATION
z −m G(z)−1 z −m

Q(z)
C(z)
D(z)
+ + R(z) + + Y (z)
R(z) + E(z) + + Y (z) F (z) G(z)
Gc (z) G(z)
− − +

G(z)
Figure 4. RC : Disturbance rejection approach.

During last years a great effort in the disturbance

rejection mechanisms has been made (Chen et al., 2016). Figure 5. RC : Youla parametrization.
In this context a RC has been proposed based on
this methodology (Chen and Tomizuka, 2014) and its It is well-known that all stabilizing controllers for a given
characteristics are shown in this section. The controller stable plant, G(z), can be written in terms of the Youla
scheme for an m-relative degree minimum phase plant, parametrization (Sánchez-Peña and Sznaier, 1998) shown
G(z), is shown in Figure 4. This system uses a baseline in Figure 5. In this case, the sensitivity and complementary
controller, Gc (z), plus a disturbance observer composed by sensitivity functions are:
the plant model and a filter, Q(z). The sensitivity function
for this closed-loop system is:
S(z) = 1 − F (z)G(z), T (z) = F (z)G(z)
1 − z −m Q(z)
S(z) = . where F (z) is a stable system. And the controller is
1 + G(z)Gc (z)
C(z) = 1−FF(z)G(z)
(z)
.
In order to transform this in a RC it is necessary to
choose Q(z) appropriately, so that the denominator of the In case of minimum-phase plants it is possible to select
internal model, I(z), appears in the numerator of S(z). F (z) = F 0 (z)G−1 (z), so the closed-loop functions become:
Consequently, Q(z) must be isolated from the equation:
1 − z −m Q(z) Ns (z) S(z) = 1 − F 0 (z), T (z) = F 0 (z).
= So (z) (1 − σW (z)H(z))
1 + G(z)Gc (z) Ds (z)
In order to impose that the controller behaves as a RC it
where Ns (z) and Ds (z) are polynomials that must be fixed is necessary to impose the appropriate shape for F 0 (z). It
(this selection requires So (z) to be stable, which can be is necessary that:
achieved by selecting and appropriate Gc (z)). So:
Ns (z)
(1 − σW (z)H(z)) Ns (z) − Ds (z) S(z) = 1 − F 0 (z) = (1 − σW (z)H(z))
Q(z) = z m . Ds (z)
Ds (z)
with Ns (z) and Ds (z) arbitrary elements. So
A simple option is choosing Ns (z) = 1 and Ds (z) = 1 −
α · σ · W (z)H(z) with |α| < 1 which generates: Ds (z) − (1 − σW (z)H(z)) Ns (z)
F 0 (z) = .
(−(1 + α) · σW (z)H(z)) Ns (z) Ds (z)
Q(z) = z m . A simple solution is Ns (z) = 1 and Ds (z) = 1 − α · σ ·
1 − α · σ · W (z)H(z)
W (z)H(z) with |α| < 1, which generates:
Although this is not the conventional shape of Q(z) in
disturbance observer based control it allows to reject (1 − α) · σ · W (z)H(z)
F 0 (z) = .
periodical signals and behave as RC. Finally the controller 1 − α · σ · W (z)H(z)
becomes: This option generates the following closed-loop transfer
Gc (z) + Q(z)z m G(z)−1 functions:
C(z) =
1 − z −m Q(z)
(1 − α) · σ · W (z)H(z)
This controller has a degree of freedom which corresponds T (z) =
to the value of α. This value plays a similar role to kr in 1 − α · σ · W (z)H(z)
the plug-in approach. 1 − σ · W (z)H(z)
S(s) = .
In case of multiplicative uncertainty the robust stability 1 − α · σ · W (z)H(z)
condition becomes: Which can be considered a generalized version of the series
kWum (z)T (z)k∞ < 1, approach (section 3).
with Finally, robust stability condition for multiplicative uncer-
1 − σW (z)H(z) tainty is :
T (z) = 1 − So (z) .
1 − α · σW (z)H(z) kWum (z)T (z)k∞ < 1.
Clearly, other selections for Ns (z) and Ds (z) are possible. Clearly, this approach can generate more complex closed-
Although these other options might provide additional loop transfer function, in particular the closed-loop poles
 d d¯

if L io
+

+ +
nonlinear
Vdc vf C vo
load
_ L _
_

d¯ d

PWM
d voltage vref
controller

Figure 6. Circuit diagram. Figure 7. Sensitivity function for RC based on plug-in and
disturbances observer approaches.
could be arbitrarily placed by choosing and appropriate
value for Ds (z). Alternatively, these degrees of freedom
could be used to optimize any criteria such as robustness.
This increase in the controller complexity would increase
the computational resources required to implement the
controller.

7. NUMERICAL EXAMPLE

The numerical example is based on the system depicted

in Figure 6. It consists of a Voltage Source Inverter
(VSI) which transforms a DC supply, Vdc , in an AC
power source, vo . The duty cycle, d, of a Pulse Width
Modulation (PWM) signal is the control action used to
switch the power transistors and the produced signal
is filtered by an LC network to obtain the output
AC voltage vo (t). The control objective is providing a
sinusoidal waveform voltage signal with specific amplitude
and frequency (vref (t)). At the same time, it is needed to
reject disturbances caused by the load. Nonlinear loads,
as the one produced by full-bridge diode rectifiers, are
of special interest since they produce disturbances with
high harmonic content. A way of measuring the system
performance is through the Total Harmonic Distortion
(THD), thus a low THD in voltage waveform (vo (t)) is
desirable.
The system transfer function is based on the LC filter
circuit equations to obtain:
Vo (s) = Gp (s)Vf (s) + Gd (s)Io (s), (3)
with
RC
Gp (s) = ,
2LCRC s2 + (2CRL RC + 2L)s + (2RL + RC )
(4)
and
2LRC + 2RL + RC
Gd (s) = − 2
, 1
2LCRC s + (2CRL RC + 2L)s + (2RL + RC )
(5)
where Gp (s) is the plant transfer function, Gd (s)Io (s) is
the disturbance signal caused by the load, L = 300 µH
and C = 80 µF are the inductive and capacitive part of
the filter, RL = 0.1 Ω and RC = 8200 Ω are the parasitic
resistance of inductance and capacitance respectively.
Figure 8. Sensitivity function for RC based on series and
Youla parametrization approaches.
Additionally, the control action is Vf (s) = (2d − 1)Vdc ,
where d ∈ [0, 1] is the duty cycle of the PWM signal
and Vdc is the DC input source. The PWM signals has
a frequency of 10KHz, and the duty cycle is updated each
period, so a sampling time of Ts = 0.0001s is used.
Although the inverter is designed to feed generic loads,
it is difficult to design a control system without some
assumptions about the concrete load being used. The load
has been modeled by means of an inductance and a resistor
in series placed in parallel with the capacitor. For this
reason, an impedance Z(s) to relate the current introduced
to the load Io (s) and the voltage supplied to it Vo (s). This
impedance is as can be seen in equations (6) and (7):
Io (s) = Z(s)Vo (s) (6)
Z(s) = (7)
L · s + RL
Taking this into account the nominal model is constructed
as:
Gp (s)
Gn (s) = .
1 + Gd (s)Z(s)
This model will be the one used to design the controller.

Figure 9. Bode diagram of 1/T (z) for all four approaches.
Figure 10. Bode diagram of 1/T (z) for all four approaches.
The goal in this type of system is feeding the load with a
sinusoidal voltage of frequency f = 50 Hz and amplitude
√ working frequency is introduced in the system. As it can
220 2 V. Consequently, the period signal of to be tracked
1
is Tp = 50 = 0.02s. Consequently, the discrete time period,
N , can be computed as N = Tp /Ts = 200. This allows to
obtain a good reconstruction of the continuous time signal.
In the following, 4 different RC systems, corresponding to
the previously introduced architectures will be designed.
In each approach different parameters have been tuned.
In the series (section 3) and the plug-in (section 4)
approaches, kr is the most relevant degree of freedom.
It is related with the most relevant closed-loop poles.
Approximately, these are the solutions of z N = 1 − kr .
These poles are homogeneously distributed over a circle
of radius |1 − kr |, which directly defines the closed-loop
settling time (Yeol et al., 2008; Garimella and Srinivasan,
1994). Due to this, the selection of kr is a trade-off between
settling time and robustness. A reasonable value is kr =
0.7.
In the disturbance rejection (section 5) and the Youla
parametrization (section 6) a parameter called α is
introduced. Similarly to kr , α is directly related with the
closed-loop poles. In fact, a relation between kr and α
can be established, α = 1 − kr , so that the closed-loop
poles from different architectures are placed in the same
location. For consistency with previous cases, α = 0.3 is
choosen.
An element present in all these architectures is the low-
pass filter. This element’s most relevant goal is introducing
robustness in the high frequency range. In this work the
following filter has been used:
0.25z 2 + 0.5z + 0.25
H(z) = , (8)
z
it has null-phase, a gain close to 1 in low and medium
frequency range and a important attenuation in the high
frequency range.
Finally, the plug-in approach (section 4) and disturbance
rejection one (section 5) contain an internal controller
in charge of providing robustness. To prevent increasing
the complexity and allowing a better comparison a
proportional controller like the following one has been
selected:
Gc (z) = 0.1. (9)
Taking this into consideration and noting that the discrete-
time plant obtained when applying the z-transform to
Gn (s) is minimum-phase plant, all the controllers intro-
duced in this work are completely defined. Figure 7 shows
the frequency response of the sensitivity function for the
plug-in approach and the disturbance observer one. As
it can be seen, both responses are similar, have small
values for the working frequency and all the harmonics.
The effect of the filter can also be seen, the small values
at harmonics increase as frequency increases. Additionally,
as the sensitivity function does not take values over 6dB
it can be stated that both of them are quite robust.
Figure 8 shows the sensitivity function for the series and
Youla architectures. As it can be seen, both schemes
provide the same result and it is quite similar to the one
observed in Figure 7.
Figure 9 shows the closed-loop time response for all
described architectures when a sinusoidal reference of the
be seen, after a small transient all closed-loop schemes
converge to the reference with null-steady state error. No
relevant differences can be observed.
Finally, the attention is turned to robustness eases. In
this type of application, the real load is not know. A way
to analyze the robustness of the control system against
changes in the load characteristics is to model the variation
in the load as a multiplicative uncertainty Wum (z) in the
plant and perform a robust stability analysis. The robust
stability condition is kWum (z)T (z)k∞ < 1, which can be
analyzed frequency by frequency as :
1
|Wum (ejωTs )| < . (10)
|T (ejωTs )|
Figure 10, shows the frequency response of T (z)−1 . As it
can be seen it is always over 0dB in all the frequency range,
and specially in the high frequency range. Consequently,
a 100% change in the impedance at each frequency can
be handled. It can be stated that the closed-loop system
is very robust. Additionally, it is possible to see that all
architectures provide a similar robustness.
 8. CONCLUSIONS
In this paper most relevant architectures used to imple-
ment RC in input-output form have been reviewed and
analyzed for the case of minimum-phase plants and its
applications to the case of an inverter has been illustrated.
As it can be shown, most architectures achieve similar re-
sults. Obtaining good steady-state results and reasonably
robustness conditions. Although it has not been shown
in this work, the plug-in approach and the disturbance
observer contains an additional degree of freedom which
can be used to slightly improve robustness and transient
behavior. Youla parametrization is the most generic by far
because it can be used to reduce all the other approaches.
Finding new approaches based on Youla parametrization
which improve transient and robustness with a limited
complexity increase is a current research area.
ACKNOWLEDGEMENTS
This work was partially supported by the spanish Min-
isterio de Educación project DPI2015-69286-C3-2-R
(MINECO/FEDER) and the catalan AGAUR project
2014 SGR 267.
REFERENCES
Ahn, H.S., Chen, Y.Q., and Moore, K.L. (2007). Iterative
Learning Control: Brief Survey and Categorization.
IEEE Transactions on Systems Man & Cybernetics Part
C, 37(6), 1099–1121.
Chen, W.H., Yang, J., Guo, L., and Li, S. (2016).
Disturbance-observer-based control and related
methods : An overview. IEEE Transactions
on Industrial Electronics, 63(2), 1083–1095. doi:
10.1109/TIE.2015.2478397.
Chen, X. and Tomizuka, M. (2014). New repeti-
tive control with improved steady-state performance
and accelerated transient. IEEE Transactions on
Control Systems Technology, 22(2), 664–675. doi:
10.1109/TCST.2013.2253102.
Chen, X. and Tomizuka, M. (2015). Overview and new
results in disturbance observer based adaptive vibration
rejection with application to advanced manufacturing.
International Journal of Adaptive Control and Signal
Processing, 29(11), 1459–1474. doi:10.1002/acs.2546.
Acs.2546.
Chen, Y., Moore, K.L., Yu, J., and Zhang, T. (2008).
Iterative learning control and repetitive control in hard
disk drive industry—a tutorial. International Journal of
Adaptive Control and Signal Processing, 22(4), 325–343.
doi:10.1002/acs.1003.
Costa-Castelló, R., Olm, J.M., Vargas, H., and Ramos.,
G.A. (2012). An educational approach to the internal
model principle for periodic signals. International
Journal of Innovative Computing, Information and
Control, 8(8), 5591–5606.
Escobar, G., Mattavelli, P., Hernandez-Gomez, M., and
Martinez-Rodriguez, P.R. (2014). Filters with linear-
phase properties for repetitive feedback. IEEE Trans-
actions on Industrial Electronics, 61(1), 405–413. doi:
10.1109/TIE.2013.2240634.
Francis, B. and Wonham, W. (1976). The internal model
principle of control theory. Automatica, 12(5), 457 – 465.
doi:10.1016/0005-1098(76)90006-6.
Garimella, S. and Srinivasan, K. (1994). Transient re-
sponse of repetitive control systems. In Proceedings
of the American Control Conference, 2909–2913. Bal-
timore.
Griñó, R. and Costa-Castelló, R. (2005). Digital repetitive
plug-in controller for odd-harmonic periodic references
and disturbances. Automatica, 41(1), 153 – 157. doi:
http://dx.doi.org/10.1016/j.automatica.2004.08.006.
Inoue, T., Nakano, M., Kubo, T., Matsumoto, S., and
Baba, H. (1982). High Accuracy control of a proton
syncrotron magnet powe supply, 3137–3142. IFAC by
Pergamon Press.
Longman, R.W. (2010). Iterative learning control and
repetitive control for engineering practice. Inter-
national Journal of Control, 73(10), 930–954. doi:
10.1080/002071700405905.
Park, S.w., Jeong, J., Yang, H.S., Park, Y.p., and Park,
N.c. (2005). Repetitive controller design for minimum
track misregistration in hard disk drives. IEEE
Transactions on Magnetics, 41(9), 2522–2528. doi:
10.1109/TMAG.2005.854338.
Ramos, G.A. and Costa-Castelló, R. (2013). Opti-
mal anti-windup synthesis for repetitive controllers.
Journal of Process Control, 23(8), 1149–1158. doi:
10.1016/j.jprocont.2013.07.004.
Ramos, G., Costa-Castelló, R., and Olm, J.M. (2013).
Digital Repetitive Control under Varying Frequency
Conditions, volume 446 of Lecture Notes in Control and
Information Sciences. Springer. ISBN: 978-3-642-37778-
5.
Sánchez-Peña, R.S. and Sznaier, M. (1998). Robust Sys-
tems Theory and Applications. Adaptive and Learning
Systems for Signal Processing, Communications and
Control Series. Wiley-Interscience.
Songschon, S. and Longman, R.W. (2003). Comparison
of the stability boundary and the frequency response
stability condition in learning and repetitive control.
International Journal of Applied Mathematics and Com-
puter Science, 13(2), 169–177.
Tomizuka, M. (1987). Zero phase error tracking algo-
rithm for digital control. Journal of Dynamic Sys-
tems, Measurement, and Control, 109(1), 65–68. doi:
10.1115/1.3143822.
Wang, L. (2016). Tutorial review on repetitive control with
anti-windup mechanisms. Annual Reviews in Control, 0,
1–14. doi:10.1016/j.arcontrol.2016.09.016.
Wang, Y., Gao, F., and Doyle, F.J. (2009). Survey on
iterative learning control, repetitive control, and run-to-
run control. Journal of Process Control, 19(10), 1589–
1600. doi:10.1016/j.jprocont.2009.09.006.
Wu, M., Xu, B., Cao, W., and She, J. (2014). Ape-
riodic disturbance rejection in repetitive-control sys-
tems. IEEE Transactions on Control Systems Technol-
ogy, 22(3), 1044–1051. doi:10.1109/TCST.2013.2272637.
Ye, Y., Tayebi, A., and Liu, X. (2009). All-pass filtering
in iterative learning control. Automatica, 45(1), 257 –
264. doi:10.1016/j.automatica.2008.07.011.
Yeol, J.W., Longman, R.W., and Ryu, Y.S. (2008). On
the settling time in repetitive control systems. In Pro-
ceedings of the 17th World Congress The International
Federation of Automatic Control Seoul, 12460–12467.
IFAC, Korea.